Back to the stories

OpenAI announces security sprint using frontier models to test and harden its systems

Score 9.5,

AI safety and security

AI is becoming part of the defensive security loop: helping find weaknesses, then helping engineers fix them.

OpenAI’s new security sprint pairs its security teams with outside researchers from Cathedral Systems. They will use OpenAI models to challenge the company’s defenses and uncover gaps.

Finding a flaw is only half the job. Engineers need to understand how it could be exploited and prepare a fix before an attacker finds it.

OpenAI says engineers used its coding tool, Codex, to investigate findings, recreate vulnerabilities in isolated test settings, and prepare patches for review. Think of it as a security rehearsal: outside researchers test the doors, and engineers use what they find to strengthen the locks.

OpenAI is also committing $1 billion in subsidised access over six months for frontline defenders protecting essential services and digital infrastructure. That could extend this defensive approach beyond OpenAI, though the announcement does not name participating operators.

The real promise is not simply finding more weaknesses. It is closing them before attackers can use them. Can this loop move faster than the threats it is meant to stop?